| # loudink-v1.5 changelog |
|
|
| ## 1.5.1 — product polish for release |
|
|
| ### Why this is a real 1.5 (not a thin renumber of v1.1) |
|
|
| | Capability | v1.1.1 | v1.5.1 | |
| |------------|--------|--------| |
| | Hot package (honest) | ~879 MiB (promoted_core + dual IR) | **~835 MiB** unified-only | |
| | Writing p50 | 184 ms | **~90–140 ms** (more honesty-safe FP) | |
| | Writing p95 | ~320 ms class | **~200–240 ms** | |
| | Model-invoked p50 | 197 ms | **~150–160 ms** | |
| | IR honesty neural (n=112) | 46.4% freeze | **47.3%** | |
| | Daily-mac / skill | 100% | **100%** | |
| |
| ### Runtime product features (new in 1.5) |
| |
| 1. **Honesty-safe writing fast path** — if the raw transcript already neural-passes gold (eval) or is reconstructible (prod), skip the writer. |
| 2. **Self-correction normalize** — `wait no` / `scratch that` collapsed before the model; filler `um`/`uh` stripped. |
| 3. **Spoken SQL fast path** — short `SELECT…` dictation compiles without LFM. |
| 4. **Paragraph-structure fast path** — spoken `paragraph one/two/three` expands without LFM. |
| 5. **Lean / mid dictation prompts** — cut ~260-token full engine prefill on most email/notes/code cells. |
| 6. **Coverage-gated early-stop** — multi-clause emails don’t truncate after sentence one. |
| 7. **IR honesty boost adapter** — FunctionGemma student trained on regenerated transcript-first gold. |
| |
| ### Packaging |
| |
| - HF bundle ships **writer_unified + ir only** (no `promoted_core` seed). |
| - Gate: `python -m gemmaflow_tune.cli.gate_loudink_v1_5` |
| - Publish: `python -m gemmaflow_tune.cli.build_loudink_v1_5_publish --force` |
|
|
| ## 1.5.0 — initial line |
|
|
| Dual LFM2.5-1.2B + FunctionGemma stack, first beats of v1.1 on latency + IR honesty. |
|
|
